Windows Agent Orchestrator — Development skill for Claude Code

Windows Agent Orchestrator (WAO)

Windows-native, headless, runtime-agnostic orchestrator for local agent runtimes (claude-code / codex / kimi-code / opencode-serve, plus an experimental DeepSeek Harness JSON-RPC adapter). Drives agents via subprocess or HTTP, records everything to JSONL transcripts, provides git worktree isolation, resume, token/cost metrics, declarative DAG workflows, and evidence-chain scorecard gating.

**Value & boundary (ADR 0018):** WAO's value is routing worker token spend onto external provider quota — it lets a Lead dispatch real work to external worker runtimes so the token bill lands on the worker's provider, instead of pulling the work back into the Lead's own context. WAO is an assisted execution control plane, not a gate and not a second semantic supervisor. WAO 自动监测，不自动监督；自动封装，不自动验收；自动呈现，不自动决策。 (English: WAO monitors, never supervises; packages, never accepts; presents, never decides.)

**New to WAO?** Start with [`AGENT_ONBOARDING.md`](AGENT_ONBOARDING.md) — it is the one authoritative path from zero to a working setup: install WAO, configure **ONE** worker, validate it, connect an MCP Host, and run a first read-only canary. You do **not** need all six runtimes or every provider credential to start.

• Minimal dependencies — plain Node ESM; only @modelcontextprotocol/sdk + zod for the MCP control surface. No Docker/WSL.
• Transcript is source of truth — every run reconstructable from runs/.jsonl.
• Windows-native — worktree isolation + process-tree cleanup tuned for Windows.

Vision

WAO is a Windows-native control plane that lets a lead agent dispatch external worker agents to do real repository work, routing the token spend onto each worker's provider quota. It keeps orchestration out of worker system prompts: workers receive normal task prompts, while WAO owns transcripts, state, isolation, workflow execution, metrics, and evidence-chain scorecard gates. Certification is advisory